bjack3 bjack3 - 7 months ago 262
Python Question

Mode of grouped data in (py)Spark

I have a spark DataFrame with multiple columns. I would like to group the rows based on one column, and then find the mode of the second column for each group. Working with a pandas DataFrame, I would do something like this:

rand_values = np.random.randint(max_value,
size=num_values).reshape((num_values/2, 2))
rand_values = pd.DataFrame(rand_values, columns=['x', 'y'])
rand_values['x'] = rand_values['x'] > max_value/2
rand_values['x'] = rand_values['x'].astype('int32')

print(rand_values)
## x y
## 0 0 0
## 1 0 4
## 2 0 1
## 3 1 1
## 4 1 2

def mode(series):
return scipy.stats.mode(series['y'])[0][0]

rand_values.groupby('x').apply(mode)
## x
## 0 4
## 1 1
## dtype: int64


Within pyspark, I am able to find the mode of a single column doing

df = sql_context.createDataFrame(rand_values)

def mode_spark(df, column):
# Group by column and count the number of occurrences
# of each x value
counts = df.groupBy(column).count()

# - Find the maximum value in the 'counts' column
# - Join with the counts dataframe to select the row
# with the maximum count
# - Select the first element of this dataframe and
# take the value in column
mode = counts.join(
counts.agg(F.max('count').alias('count')),
on='count'
).limit(1).select(column)

return mode.first()[column]

mode_spark(df, 'x')
## 1
mode_spark(df, 'y')
## 1


I'm at a loss for how to apply that function to grouped data. If it's not possible to directly apply this logic to a DataFrame, is it possible to achieve the same effect by some other means?

Thank you in advance!

Answer

Solution suggested by zero323.

Original solution: http://stackoverflow.com/a/35226857/1560062

First, count the occurances of each (x, y) combination.

counts = df.groupBy(['x', 'y']).count().alias('counts')
counts.show()
## +---+---+-----+
## |  x|  y|count|
## +---+---+-----+
## |  0|  1|    2|
## |  0|  3|    2|
## |  0|  4|    2|
## |  1|  1|    3|
## |  1|  3|    1|
## +---+---+-----+

Solution 1: Group by 'x', aggregate by taking the maximum value of the counts in each group. Finally, Drop the 'count' column.

result = (counts
          .groupBy('x')
          .agg(F.max(F.struct(F.col('count'),
                              F.col('y'))).alias('max'))
          .select(F.col('x'), F.col('max.y'))
         )
result.show()
## +---+---+
## |  x|  y|
## +---+---+
## |  0|  4|
## |  1|  1|
## +---+---+

Solution 2: Using a window, partition by 'x', and order by the 'count' column. Now, pick the first row in each of the partitions.

win = Window().partitionBy('x').orderBy(F.col('count').desc())
result = (counts
          .withColumn('row_num', F.rowNumber().over(win))
          .where(F.col('row_num') == 1)
          .select('x', 'y')
         )
result.show()
## +---+---+
## |  x|  y|
## +---+---+
## |  0|  1|
## |  1|  1|
## +---+---+

The two results have a different outcome because of the way the rows are sorted. If there are no ties, the two methods give the same result.

Comments